My latest build of the 4.0 branch is generating a lot of remarks, though I'm not explicitly asking for them. llvm/4.0/bin/llc /tmp/5-oROzlY7V8.llvm -mcpu=native -O2 -code-model=large -o /tmp/X-ot2IA_JlK.s remark: foobar.cpp:1150:1: failed to move load with loop-invariant address because the loop may invalidate its value …
I'll try to reproduce.
OK, that's probably because PPC schedules an IR LICM pass unlike other targets?
OK, reproduced: $ ./bin/llc ../test/Transforms/LICM/opt-remarks-intervening-store.ll -O2 -march=ppc64 remark: /tmp/kk.c:2:20: failed to move load with loop-invariant address because the loop may invalidate its value
Created attachment 18068 [details] Fix The attached patch fixes it for me.
Hans, can you please commit this to the 4.0 branch? This is a partial backport of r293110 from trunk.
I suppose this regressed with r291648? We're very late in the release process and I was hoping to promote rc4 to final today. I would rather not do an rc5. How bad is this? Could it be emitting remarks for Clang even if users haven't asked for them?
(In reply to Hans Wennborg from comment #6) > How bad is this? Could it be emitting remarks for Clang even if users > haven't asked for them? Oh wait, your patch is for llc. I don't think we can hold the release for this.
Sure. Can you still commit it on the branch for me?
(In reply to Hans Wennborg from comment #7) > (In reply to Hans Wennborg from comment #6) > > How bad is this? Could it be emitting remarks for Clang even if users > > haven't asked for them? > > Oh wait, your patch is for llc. > > I don't think we can hold the release for this. I agree, we shouldn't hold the release. It should end up in the release branch for the point release, however.
(In reply to Hal Finkel from comment #9) > (In reply to Hans Wennborg from comment #7) > > (In reply to Hans Wennborg from comment #6) > > > How bad is this? Could it be emitting remarks for Clang even if users > > > haven't asked for them? > > > > Oh wait, your patch is for llc. > > > > I don't think we can hold the release for this. > > I agree, we shouldn't hold the release. It should end up in the release > branch for the point release, however. Yes. Please file a blocker against 32061 for that.
Done.
Eric, can you please confirm if this fixes the problem for you too?
(In reply to Adam Nemet from comment #12) > Eric, can you please confirm if this fixes the problem for you too? Hi Adam, can you tell me what commit to look for? Thanks in advance.
Eric, please apply the patch attached to this bug. It's not yet committed I think.
The patch looks good. Not sure about "resolved" status though, since it does not appear to be checked in.
Thanks. Good point on the status, moved it back. When Hans checks it in we can move it back to RESOLVED.
Description of the bug for the 4.0.1 release: Partial merge of r293110 into the 4.0 branch: Avoid printing optimization remarks from llc unless they are enabled.
Merged: r299409